some doubts about the forward loss compute

[cs-heibao]

here the C-IOU loss computation also need the positive index? in my understanding, if we still consider positive sample during c-iou loss compute, why the paper said this kind of loss can improve the conditions: anchors with none overlaps with gt compared with IOU-Loss and G-IOU Loss, thanks if self.loss == 'SmoothL1': loss_l = F.smooth_l1_loss(loc_p, loc_t, reduction='sum') else: giou_priors = priors.data.unsqueeze(0).expand_as(loc_data) loss_l = self.gious(loc_p,loc_t,giou_priors[pos_idx].view(-1, 4))

[Zzh-tju]

As we know, convolution is a sequence of local operations. Its receptive field is limited. During bbox regression, we cannot learn the box that is too far away. If we force models to learn distant objects, it will damage the network. So in practice, it is necessary to select positive samples. And this is the reason that there is limited room for improvement of all these bbox regression losses.

[cs-heibao]

@Zzh-tju So, IOU-Loss、G-IOU Loss and D-IOU( or C-IOU)Loss all only works for anchors which exist overlap with gt through different ways. The paper D-IOU Loss specially consider other three directions(overlap,central point distance and aspect ratio), and then accelerate the bbox regression with better results as well?

[Zzh-tju]

yes, that is CIoU loss

[feixiangdekaka]

IOU 51.01
LGIoU 51.06

LDIoU 51.31

LCIoU 51.44 The improvement for SSD is relatively small, 0.43 point.